Quintin Pope: AI discourse overheating as nanochat, modded-nanogpt benchmarks stall

QuintinPope5 · x · 2026-09-29

Researcher Quintin Pope argues the current AI memetic environment is starting to overheat in takeoff-speed debates. He maintains AI training already has a recursive self-improvement aspect, since NN inductive biases align with the target function. He also notes the 'fast LLM training' benchmarks nanochat and modded-nanogpt have stalled in recent months despite Karpathy's autoresearch effort.

Related event: Debating default alignment: does RL post-training twist model minds(9 posts)→

Original post →

More from AGI Musings

AGI Musings channel →